Great post! I don't have much to add, except for one nitpick about a footnote:

In the orginal AlphaGo Zero design, black always gets the all-one feature and therefore some information about the size of the board, while white always gets the all-zero feature and no information; white still infers the board size, of course, by the fact that there are never stones placed outside of the board.

It's even easier to infer the board size: just add a bias of 1 and an all-zero filter in the first convolution layer and you can immediately create your own all-one feature plane.